raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

gpu: reflect tessellation strides from memory address dataflow

Browse Source
This commit is contained in:
nmzik
2026-09-13 08:58:40 +04:00
parent 7516068662
commit 96d62b3367
2 changed files with 203 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ShaderRecompiler.cpp
+143 -43
View File
@@ -511,53 +511,153 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
return result;
}
struct TessellationAddress {
enum class Kind { Unknown, Affine, PackedControlPoint };
Kind kind = Kind::Unknown;
uint32_t coefficient = 0;
uint32_t constant = 0;
};
uint32_t ReflectTessellationStride(const Decoder::Program& program, bool local,
uint32_t control_points, uint32_t input_stride) {
using namespace Decoder;
using Address = TessellationAddress;
using Kind = Address::Kind;
std::array<Address, IR::NumVectorRegs> registers {};
registers[local ? 3u : 1u] =
local ? Address {Kind::Affine, 1u, 0u} : Address {Kind::PackedControlPoint};
const auto constant = [](uint32_t value) { return Address {Kind::Affine, 0u, value}; };
const auto read = [&](const Operand& operand) -> Address {
if (operand.absolute || operand.negate || operand.sdwa_sext || operand.op_sel ||
operand.op_sel_hi || operand.negate_hi || operand.dpp)
return {};
Address value;
if (operand.kind == OperandKind::IntegerInlineConstant ||
operand.kind == OperandKind::LiteralConstant) {
value = constant(operand.value);
} else if (operand.kind == OperandKind::Vgpr) {
value = registers.at(operand.reg);
}
// The native packed HS ID's low byte is the patch ordinal. The logical
// interface addresses one patch, as in the frontend's v1 initialization.
if (value.kind == Kind::PackedControlPoint && operand.sdwa_sel == 0u) return constant(0u);
return operand.sdwa_sel == 6u ? value : Address {};
};
const auto affine = [](uint64_t coefficient, uint64_t offset) {
return coefficient <= UINT32_MAX && offset <= UINT32_MAX
? Address {Kind::Affine, static_cast<uint32_t>(coefficient),
static_cast<uint32_t>(offset)}
: Address {};
};
const auto add = [&](Address lhs, Address rhs) {
return lhs.kind == Kind::Affine && rhs.kind == Kind::Affine
? affine(uint64_t {lhs.coefficient} + rhs.coefficient,
uint64_t {lhs.constant} + rhs.constant)
: Address {};
};
const auto multiply = [&](Address lhs, Address rhs) {
if ((lhs.kind == Kind::Affine && lhs.coefficient == 0u && lhs.constant == 0u) ||
(rhs.kind == Kind::Affine && rhs.coefficient == 0u && rhs.constant == 0u))
return constant(0u);
if (lhs.kind != Kind::Affine || rhs.kind != Kind::Affine) return Address {};
if (rhs.coefficient != 0u) std::swap(lhs, rhs);
return rhs.coefficient == 0u ? affine(uint64_t {lhs.coefficient} * rhs.constant,
uint64_t {lhs.constant} * rhs.constant)
: Address {};
};
const auto low24 = [&](Address value) {
if (value.kind == Kind::Affine && value.coefficient == 0u)
return constant(value.constant & 0xffffffu);
return value.kind == Kind::Affine &&
uint64_t {value.coefficient} * (control_points - 1u) + value.constant <=
0xffffffu
? value
: Address {};
};
uint32_t stride = 0;
for (const auto& inst: program.instructions) {
// Stage exits preserve the active path's definitions. An internal join
// would require merging register definitions.
EXIT_NOT_IMPLEMENTED(IsDirectBranch(inst.opcode) &&
inst.branch_target != program.instructions.back().pc);
const bool local_store =
inst.opcode == Opcode::DS_WRITE_B32 || inst.opcode == Opcode::DS_WRITE2_B32;
const bool buffer_store = inst.opcode == Opcode::BUFFER_STORE_DWORD ||
inst.opcode == Opcode::BUFFER_STORE_DWORDX2 ||
inst.opcode == Opcode::BUFFER_STORE_DWORDX3 ||
inst.opcode == Opcode::BUFFER_STORE_DWORDX4;
const bool control_store =
!local && buffer_store && inst.src2.kind == OperandKind::Sgpr && inst.src2.reg == 2u;
const bool control_read =
!local && (inst.opcode == Opcode::DS_READ_B32 || inst.opcode == Opcode::DS_READ2_B32);
if ((local && local_store) || control_store || control_read) {
const auto address = read(inst.src0);
if (address.kind != Kind::Affine) {
EXIT("%s tessellation address is not affine at pc 0x%08x\n", local ? "LS" : "HS",
inst.pc);
}
if (address.coefficient != 0u) {
const auto expected = control_read ? input_stride : stride;
EXIT_NOT_IMPLEMENTED((address.coefficient & 3u) != 0u ||
(expected != 0u && expected != address.coefficient));
if (!control_read) stride = address.coefficient;
}
}
// Store vdata is a source, despite occupying the decoder's dst field.
if (local_store || buffer_store || inst.dst.kind != OperandKind::Vgpr) continue;
const auto lhs = read(inst.src0);
const auto rhs = read(inst.src1);
const auto third = read(inst.src2);
Address value;
switch (inst.opcode) {
case Opcode::V_BFE_U32:
if (lhs.kind == Kind::PackedControlPoint && rhs.kind == Kind::Affine &&
rhs.coefficient == 0u && third.kind == Kind::Affine &&
third.coefficient == 0u) {
if (rhs.constant == 8u && third.constant == 5u) value = {Kind::Affine, 1u, 0u};
if (rhs.constant == 0u && third.constant == 8u) value = constant(0u);
}
break;
case Opcode::V_MUL_U32_U24: value = multiply(low24(lhs), low24(rhs)); break;
case Opcode::V_MAD_U32_U24: value = add(multiply(low24(lhs), low24(rhs)), third); break;
case Opcode::V_LSHL_ADD_U32:
if (rhs.kind == Kind::Affine && rhs.coefficient == 0u && rhs.constant < 32u)
value = add(multiply(lhs, constant(1u << rhs.constant)), third);
break;
case Opcode::V_LSHLREV_B32:
if (lhs.kind == Kind::Affine && lhs.coefficient == 0u && lhs.constant < 32u)
value = multiply(rhs, constant(1u << lhs.constant));
break;
case Opcode::V_SUB_NC_U32:
if (lhs.kind == Kind::Affine && rhs.kind == Kind::Affine &&
lhs.coefficient >= rhs.coefficient && lhs.constant >= rhs.constant)
value = affine(lhs.coefficient - rhs.coefficient, lhs.constant - rhs.constant);
break;
default: break;
}
if (inst.dst.sdwa_sel != 6u || inst.dst.op_sel || inst.dst.omod || inst.dst.clamp ||
inst.dst.dpp)
value = {};
for (uint32_t index = 0;
index < DecodedDstSize(inst) && inst.dst.reg + index < registers.size(); index++) {
registers[inst.dst.reg + index] = {};
}
registers.at(inst.dst.reg) = value;
}
EXIT_NOT_IMPLEMENTED(stride == 0u);
return stride;
}
} // namespace
void AnalyzeTessellationPrograms(std::span<const uint32_t> local, std::span<const uint32_t> control,
ShaderTessellationInputInfo& info) {
using namespace Decoder;
const auto constant = [](const Operand& operand, uint32_t value) {
return (operand.kind == OperandKind::IntegerInlineConstant ||
operand.kind == OperandKind::LiteralConstant) &&
operand.value == value;
};
auto decoded = DecodeFrontProgram(local);
uint32_t local_address = UINT32_MAX;
for (const auto& inst: decoded.instructions) {
if (inst.opcode == Opcode::V_MUL_U32_U24 && inst.src1.kind == OperandKind::Vgpr &&
inst.src1.reg == 3u &&
(inst.src0.kind == OperandKind::IntegerInlineConstant ||
inst.src0.kind == OperandKind::LiteralConstant)) {
EXIT_IF(info.ls_stride != 0u && info.ls_stride != inst.src0.value);
info.ls_stride = inst.src0.value;
local_address = inst.dst.reg;
}
if (inst.family == Family::DS) {
EXIT_NOT_IMPLEMENTED(inst.src0.kind != OperandKind::Vgpr ||
inst.src0.reg != local_address);
}
}
decoded = {};
DecodeProgram(control, decoded);
uint32_t control_point = UINT32_MAX;
uint32_t control_address = UINT32_MAX;
for (const auto& inst: decoded.instructions) {
if (inst.opcode == Opcode::V_BFE_U32 && inst.src0.kind == OperandKind::Vgpr &&
inst.src0.reg == 1u && constant(inst.src1, 8u) && constant(inst.src2, 5u)) {
control_point = inst.dst.reg;
}
if (inst.opcode == Opcode::V_LSHL_ADD_U32 && inst.src0.kind == OperandKind::Vgpr &&
inst.src0.reg == control_point &&
(inst.src1.kind == OperandKind::IntegerInlineConstant ||
inst.src1.kind == OperandKind::LiteralConstant) &&
inst.src1.value < 32u) {
info.hs_stride = 1u << inst.src1.value;
control_address = inst.dst.reg;
}
}
EXIT_NOT_IMPLEMENTED(local_address == UINT32_MAX || control_address == UINT32_MAX ||
info.ls_stride == 0u || info.hs_stride == 0u ||
(info.ls_stride & 3u) != 0u || (info.hs_stride & 3u) != 0u);
const auto local_program = DecodeFrontProgram(local);
Decoder::Program control_program;
Decoder::DecodeProgram(control, control_program);
info.ls_stride = ReflectTessellationStride(local_program, true, info.input_control_points, 0u);
info.hs_stride = ReflectTessellationStride(control_program, false, info.output_control_points,
info.ls_stride);
LOGF("Tessellation interface: input_cp=%u output_cp=%u ls_stride=%u hs_stride=%u\n",
info.input_control_points, info.output_control_points, info.ls_stride, info.hs_stride);
}